Float Point Precision about redisjson HOT 6 CLOSED

@dvirsky I agree. The flag for serialization seems to be wrong at this line:

https://github.com/RedisLabsModules/rejson/blob/master/src/json_object.c#L342

I did some experiment for the flags:

(check out this snippet: https://repl.it/GjP1/0)

#include <stdio.h>
#include <math.h>
#include <float.h>

int main ()
{
   float a = 1.333333;
  
   printf("Should be: %lf, but losing precision: %g", a, a);
   
   return(0);
}

Output of the above is:

Should be: 1.333333, but losing precision: 1.33333 

Supposedly %g outputs the shorter of %e or %f, which in this case it uses %f and loses precision. See a similar issue here:

json-c/json-c#59

I get the need to save space for serialization, but in this case it should either do more detection in terms of precision or use %f outright.

from redisjson.

@itamarhaber I think it's a serialization issue, internally we shouldn't be losing precision

from redisjson.

And changing %e to %lf fixes the problem:

127.0.0.1:6380[7]> JSON.GET world
"1.333333"

Not sure about the implication in terms of the memory footprint, however.

from redisjson.

It shouldn't matter in terms of memory since this is only done when serializing the objects. Maybe make a PR and we'll continue the discussion over there?

from redisjson.

Here you go.

from redisjson.

Apparently, %.17g is the more correct version:

http://stackoverflow.com/questions/16839658/printf-width-specifier-to-maintain-precision-of-floating-point-value

from redisjson.